Speech processing techniques are disclosed that enable determining a text representation of alphanumeric sequences in captured audio data. Various implementations include determining a contextual biasing finite state transducer (FST) based on contextual information corresponding to the captured audio data. Additional or alternative implementations include modifying probabilities of one or more candidate recognitions of the alphanumeric sequence using the contextual biasing FST, where the FST further comprises a grammar as well as a speller finite state transducer.
Legal claims defining the scope of protection. Each claim is shown in both the original legal language and a plain English translation.
2. The method of claim 1, wherein the ASR model is a recurrent neural network transducer (RNN-T) model, and wherein the ASR engine further comprises a beam search portion.
This method for automatic speech recognition (ASR) involves determining a text representation of alphanumeric sequences from captured audio data. The ASR engine used in this process specifically employs a Recurrent Neural Network Transducer (RNN-T) model. Furthermore, the ASR engine incorporates a beam search algorithm. This system also applies a contextual biasing Finite State Transducer (FST), which includes both a grammar and a speller FST, to modify the probabilities of various candidate recognitions for the alphanumeric sequences to improve accuracy.
8. The method of claim 1, wherein the audio data is captured via one or more microphones of a client device.
This method for automatic speech recognition (ASR) determines a text representation of alphanumeric sequences from captured audio data. A key aspect is that the audio data itself is captured through one or more microphones that are part of a client device. An ASR engine processes this audio data, and a contextual biasing Finite State Transducer (FST), comprising a grammar and a speller FST, is applied to modify the probabilities of candidate recognitions for the alphanumeric sequences to enhance recognition accuracy.
10. The method of claim 1, wherein the alphanumeric sequence includes at least one number and includes at least one letter.
This method performs automatic speech recognition (ASR) to determine text representations of alphanumeric sequences found within captured audio data. A defining characteristic is that these alphanumeric sequences include at least one number and at least one letter (e.g., "A1" or "R2D2"). An ASR engine processes the audio data, and to improve recognition accuracy for these mixed alphanumeric sequences, a contextual biasing Finite State Transducer (FST) that includes both a grammar and a speller FST is used to modify the probabilities of candidate recognitions.
11. The method of claim 1, wherein the ASR model portion of the ASR engine is an end-to-end speech recognition model.
This method for automatic speech recognition (ASR) determines a text representation of alphanumeric sequences from captured audio data. The ASR engine responsible for this task specifically utilizes an end-to-end speech recognition model. To further refine the recognition process, especially for alphanumeric sequences, a contextual biasing Finite State Transducer (FST) that incorporates both a grammar and a speller FST is applied to modify the probabilities of candidate recognitions.
12. The method of claim 1, wherein the ASR engine is trained using a set of training instances, and wherein the alphanumeric sequence is not in the set of training instances.
This method uses an automatic speech recognition (ASR) engine to determine text representations of alphanumeric sequences from captured audio data. A significant aspect is that while the ASR engine was trained using a specific set of training instances, the particular alphanumeric sequence being recognized in the current audio data was not included in that initial training set. Despite this lack of direct training data, a contextual biasing Finite State Transducer (FST), comprising a grammar and a speller FST, is still applied to modify candidate recognition probabilities for this out-of-training-set alphanumeric sequence.
13. The method of claim 1, wherein the ASR engine is trained using a set of training instances, and wherein the alphanumeric sequence occurs a number of times, in the set of training instances, that is below a threshold value.
This method employs an automatic speech recognition (ASR) engine to determine text representations of alphanumeric sequences from captured audio data. The ASR engine was trained using a specific set of training instances, and the particular alphanumeric sequence being recognized appears in this training data, but only a limited number of times, below a predefined threshold value. To enhance the recognition accuracy for such rarely encountered sequences, a contextual biasing Finite State Transducer (FST), which includes both a grammar and a speller FST, is applied to modify the probabilities of candidate recognitions.
16. The computing system of claim 15, wherein the ASR model is a recurrent neural network transducer (RNN-T) model, and wherein the ASR engine further comprises a beam search portion.
This computing system is designed for automatic speech recognition (ASR) of alphanumeric sequences within captured audio data. It includes a processor, memory, and an ASR engine. The ASR engine's underlying model is specifically a Recurrent Neural Network Transducer (RNN-T) and also incorporates a beam search component for processing. The system is configured to apply a contextual biasing Finite State Transducer (FST), which includes both a grammar and a speller FST, to modify the probabilities of candidate recognitions for the identified alphanumeric sequences, thereby accurately determining their text representation.
Cooperative Patent Classification codes for this invention. Click any code to explore related patents in that topic.
January 17, 2020
March 26, 2024
Browse 5M+ US patents with plain-English claim translations and AI-generated analysis.